EFR Analysis

Recorded EEG was filtered using MATLAB's (R2020b, MathWorks, Natick, MA) zero-phase filtfilt function to apply a 16th order infinite-impulse-response Butterworth bandpass filter between 65 and 300 Hz for removing noise outside the frequency range containing the EFRs at f0 and neighboring noise tracks. EEG was then considered in sequential segments of 1.0445 seconds called epochs for the purpose of identifying and removing periods of excessive noise likely of myogenic origin. There were eight epochs per trial and a noise metric (NM) was calculated for each epoch as the mean spectral amplitude between 65 and 300 Hz. A rejection threshold using all NMs was determined as the third NM quartile + 1.5 times the interquartile range. Epochs for which NMs exceed this rejection threshold were excluded from calculation of an average trial using all remaining epochs. A 10-ms correction was applied to undo estimated physiological processing delay and thus better align the EEG of the average trial with the f0 track calculated for each vowel and described below. This delay correction is deemed reasonable for response frequencies greater than 80 Hz. ^{10 36 37 38} Using the analysis methods described here with relatively long duration vowels and relatively flat f0 contours, EFR amplitude estimates are not very sensitive to the exact delay correction applied. ³⁹ The final preprocessing step was to average the two halves of the average trial. Responses within each half are elicited from different stimulus polarities so this step averages the response across polarities ³⁵ to create a final average for Fourier analysis that potentially contains an EFR for each of the 10 vowels.

EFR properties for a participant were estimated with a software-implemented Fourier analyzer. ^{13 18} A single amplitude and phase estimate was calculated for the EFR elicited by each vowel stimulus using the analysis window durations given in the bottom right of Table 1 . Vowel f0 tracks were estimated with Praat and f0 reference cosine and sine sinusoids were calculated and multiplied with the delay-corrected average EEG. Across each analysis window, mean values of these products were independently obtained for real and imaginary components and subsequently used to calculate EFR amplitude and phase. Neighboring noise was estimated using 10 parallel noise tracks separated from the f0 track by integer multiples of the reciprocal of analysis window duration. To avoid including 60 Hz in noise estimates for male vowels with the lowest f0, only two noise tracks were below f0 and eight above. The same arrangement of noise tracks was used for all vowels to maintain consistency.